Viz.ai Stroke AI — FDA Device Record: K192872

Device Identification

Intended Use (Authorized Scope)

Viz LVO is authorized as a software-only device intended to analyze CT angiography (CTA) images for findings consistent with large vessel occlusion. The device automatically notifies a specialist — typically a neurointerventionalist or stroke neurologist — when a suspected LVO is detected, enabling faster mobilization of the stroke team.

The device is intended to assist in the detection of large vessel occlusions (LVOs) in the anterior circulation of the brain on computed tomography angiography (CTA) images. The software analyzes CTA images and automatically notifies a specialist when a suspected LVO is detected. The device is not intended to replace the clinical judgment of a physician.

FDA 510(k) Summary, K192872, cleared February 7, 2020

Regulatory Pathway: Why 510(k)?

Viz.ai pursued the 510(k) premarket notification route, demonstrating substantial equivalence to a predicate device — Aidoc Medical's K183322, which had received clearance in 2018 for intracranial hemorrhage detection. The 510(k) pathway does not require the same level of prospective clinical trial evidence as a PMA, but does require demonstrating that the device's technological characteristics and intended use are substantially equivalent to the predicate.

This matters for practitioners evaluating the device: 510(k) clearance means the FDA has determined the device is as safe and effective as a legally marketed predicate — not that it has been independently validated in a prospective randomized trial. Post-clearance real-world evidence is what fills that gap, and it has accumulated for Viz LVO more than for most cleared stroke AI tools.

Clinical Context: What the Device Actually Does

Acute ischemic stroke from large vessel occlusion is time-critical. Every minute of delayed thrombectomy translates to roughly 1.9 million neurons lost. The clinical bottleneck is often not imaging acquisition — hospitals increasingly have fast CT protocols — but the time between imaging completion and specialist notification.

Viz LVO sits in that gap. When a CTA is acquired and transferred to the PACS, the software processes the images automatically and pushes a mobile alert to the on-call neurointerventionalist or stroke team if an LVO is detected. The radiologist still reads the study; Viz LVO runs in parallel, not as a replacement read.

The workflow integration model is important to understand. Viz.ai operates as a cloud-based platform: CTA DICOM data is routed to Viz servers, processed, and results are returned via a mobile application. This means the device's latency depends on network transfer speed and cloud processing time, not just algorithm performance. Institutions with slower PACS-to-cloud transfer pipelines will see longer notification delays regardless of the algorithm's intrinsic speed.

Performance Data from the Submission Record

The 510(k) submission included a retrospective reader study comparing Viz LVO detection against ground truth established by expert consensus reads. The dataset covered a multi-site cohort of CTA studies.

Post-Market Evidence: What Has Been Published

Unlike many FDA-cleared AI devices that have minimal post-clearance peer-reviewed literature, Viz LVO has accumulated a meaningful body of published evidence — though it remains predominantly retrospective and single-center or limited multi-site in design.

Time-to-Notification Studies

Several published studies have examined door-to-notification time and door-to-puncture intervals before and after Viz LVO deployment. A study published in the Journal of NeuroInterventional Surgery (2021) found a statistically significant reduction in door-to-puncture time at a comprehensive stroke center after implementing the platform — median reduction reported at approximately 30 minutes. However, these pre-post comparisons carry confounding risks: stroke team composition, protocol changes, and case volume shifts can all independently affect these intervals.

Detection Performance in Real-World Cohorts

Post-market detection performance studies have generally confirmed sensitivity in the high-80s to low-90s percent range for anterior circulation M1 and internal carotid artery occlusions. Performance is lower for M2 segment occlusions — a known limitation across LVO detection algorithms, not specific to Viz.ai. M2 occlusions are smaller, more variable in anatomy, and less reliably detected by current deep learning approaches.

False Positive Rate Considerations

False positives — alerts triggered in patients without LVO — are the primary operational friction point in deployed systems. Each false positive requires the on-call specialist to review the case, which adds burden during off-hours and can erode trust in the alert system over time. Published institutional reports have noted false positive rates ranging from 5% to 15% depending on scanner protocol and patient population characteristics. Institutions with higher proportions of patients with dense calcified vessels or prior stroke changes tend to see higher false positive rates.

Known Limitations and Scope Boundaries

• Anterior circulation only: Basilar artery occlusion and other posterior circulation LVOs are outside the cleared scope and not reliably detected.
• CTA dependency: The device requires CTA input. Non-contrast CT, CT perfusion, or MRI sequences are not processed by this cleared version.
• Cloud latency: Processing time includes data transfer to cloud infrastructure. In practice, total notification time from scan completion ranges from 2 to 6 minutes depending on network conditions.
• Not a replacement read: The device output is a notification trigger, not a diagnostic report. A qualified radiologist or physician must still interpret the CTA.
• Image quality dependence: Motion artifact, poor contrast bolus timing, or suboptimal CTA protocol can degrade detection performance. The algorithm does not flag poor-quality inputs.
• Population representativeness: Training data composition is not publicly disclosed in the submission summary. Performance in populations underrepresented in training data (e.g., pediatric patients, patients with unusual vascular anatomy) is not characterized.

Equity and Population Considerations

The 510(k) submission summary does not disclose demographic breakdown of the training or validation dataset. This is a gap common to most AI device submissions from this period — FDA guidance on algorithmic transparency and demographic subgroup reporting has strengthened in subsequent years, but earlier submissions predating the 2021 AI/ML action plan were not required to report this level of detail.

Stroke incidence and outcomes are not demographically uniform. Black patients in the United States experience higher stroke incidence and worse outcomes than white patients on average, and there are documented differences in vascular anatomy distribution and comorbidity burden that could affect LVO detection performance. Whether Viz LVO's detection performance is equivalent across racial and ethnic subgroups has not been publicly characterized in peer-reviewed literature as of this record's review date.

Subsequent Clearances and Platform Expansion

K192872 covers the Viz LVO product specifically. Viz.ai has received additional 510(k) clearances for related products, including Viz ICH (intracranial hemorrhage detection) and Viz PE (pulmonary embolism). These are separate submissions with separate authorized scopes — the LVO clearance does not extend to those indications. Practitioners evaluating the broader Viz.ai platform should verify the specific submission number and intended use for each indication being considered.

Primary Source and Verification

This record is derived from the FDA 510(k) premarket notification database. The authoritative source for all regulatory details is the FDA's official submission record, accessible via the FDA 510(k) database using submission number K192872. The decision summary and 510(k) summary document are publicly available and contain the full intended use statement, predicate comparison, and performance testing description.